Learning to rank (L2R) works by constructing a ranking model from training data so that, given unseen data (query), a somewhat similar ranking is produced. Almost all work in L2R focuses on ranking accuracy leaving performance and scalability overlooked. In this work we present a fast and scalable manycore (GPU) implementation for an on-demand L2R technique that builds ranking models on the fly. Our experiments show that we are able to process a query (build a model and rank) in only a few milliseconds, achieving a speedup of 508x over a serial baseline and 4x over a parallel baseline for the best case. We extend the implementation to work with multiple GPUs, further increasing the speedup over the parallel baseline to approximately x16 when using 4 GPUs.